CoreWeave Inc. has introduced CoreWeave Forge, a platform designed to integrate model deployment, evaluation, and improvement. The new service includes a reinforcement learning feature called Rollouts, which is currently available in preview. This tool facilitates repeated cycles of generating training responses and updating models.
To optimize GPU utilization, CoreWeave has implemented a weight synchronization method that allows for hot starts, pulling weights from nearby peers rather than relying solely on object storage. Additionally, the company’s AI Object Storage now supports cross-region writes, enabling faster data access for post-training jobs.
You.com Inc. has joined the CoreWeave partner network to provide search capabilities for AI agents. By utilizing the RL Rollouts feature, the companies successfully completed the post-training of the Nemotron 3.5 Lightning model in eight hours. This process integrated You.com’s web search tools and was conducted in collaboration with Nvidia.
